The invention relates to the information image recognition and can be used for adaptive processing and classification of information flow.
At present, a variety of adaptive image recognition techniques are known for both pictures and other forms of images such as sound or tactile ones.
In the paper xe2x80x9cResearch decision-making system by complex-image fragments using neuron-like algorithmsxe2x80x9d (Izv. Vyssh. Uchebn. Zaved., Radiofiz., vol. 37, 1994, No. 8, pp. 961-986 (in Russian)) by V. G. Yakhno et al. is described an approach which outlines the following ways for solving the problem of information image adaptive recognition: signal encoding; back recovering of signals from their code descriptions; forming a set of estimates for code descriptions and recovered signals; forming a tree of possible solutions; determining the efficiency or correctness of encoding; recovering and estimating; and realigning operating algorithms in the case of their inefficiency.
In this approach, as mentioned above, only general ways are outlined, but there are no particular operation sequence for adaptive image recognition and no devices for implementing that approach. The recognition algorithms were assumed to be predetermined in that paper.
In the RF Patent No. 2090928 (Int. C1. G 06 K 9/00, Sep. 20, 1997), is described a method for adaptive signal recognition on the state of object, which method includes steps of creating recognition standards and correcting them during a training, comparing these standards with a signal being analyzing, forming estimates of object states, and comparing these estimates with the predetermined recognition threshold, the step of correcting the formed standards being iterative. A drawback of this method is that the recognition is performed by comparing the signal being recognized with the standards obtained at the training stage and not liable to correction during recognition, i.e., obtained without checking the degree of actual correspondence between these standards and the signal being recognized. Thus, a situation can arise in which the signal being recognized is different from the signals by which the standards were formed and corrected at the training stage, which results in recognition errors.
The U.S. Pat. No. 4,491,962 (Int. C1. G 06 K 9/00, Jan. 1, 1985) describes a recognition method which determines the optimal threshold for obtaining a binary image signal. In this method, the video signal of an object is converted to a binary image signal on the basis of each of various values of the threshold level, in which determining a template expected to lie somewhere on the object surface is chosen as a standard one. Then a set of partial templates is cut out of the given region of a binary image signal corresponding to each value of the threshold level, and the degree of coincidence between each partial template and the indicated standard template is determined. The value of the threshold level is chosen which corresponds, as a whole, to the maximum degree of coincidence between this set of partial templates and the standard template taking into account the coincidence degrees for each value of the threshold levels. The object image is divided into parts, and the maximum degree of coincidence between the set of partial templates and the standard template is chosen. This process is performed for all parts of the image, and the final sum of maximum values of all surface parts is used as the coincidence degree in whole. A drawback of this method is that the image which can be absent in the picture being recognized and is not liable to correction can be used as the standard template. This can lead to incorrect recognition.
This method is implemented in a system for adaptive recognition of information images described in the same U.S. Pat. No. 4,491,962 and comprising an information image adaptive processing unit intended for processing an, input information image in accordance with the algorithm developed beforehand for recognizing information images; a memory unit intended for storing the information image processed by the information image adaptive processing unit at least until finishing the image recognition procedure; a comparison unit intended for comparing the input information image being processed in the information image adaptive processing unit with the model image and for estimating discrepancies between the compared images; and a decision unit intended for deciding on the recognition of the information image. The drawbacks of this system are the same as those of the method being implemented by this system.
The U.S. Pat. No. 5,054,093 (Int. C1. G 06 K 9/00, Oct. 1, 1991) describes a parallel multiunit adaptive nonlinear coupler and classifier for implementing another method for an adaptive recognition which is closest to the claimed method. This known method comprises steps of: forming in advance models of probable information images (pattern prototypes related to different classes of patterns); processing each input information image in accordance with the algorithm developed in advance for recognizing information images; comparing the input information image being processed (input pattern) with the prototype; and correcting the recognition algorithm by storing a new prototype if the input pattern does not coincide with a region of the initial prototype; and, in the case of coincidence, modifying the prototype in the memory for matching it with the input pattern. A drawback of this method is that only one recognition algorithm with replaceable prototypes is used, which narrows the applicability of this method. Moreover, in this method, a coincidence is estimated with the prototypes of the information images already processed, which precludes to check the correctness of the processing procedure itself and its effect in the result being obtained.
As was mentioned above, this method is implemented in an apparatus which is chosen as the nearest analogue of the claimed one and comprises: an information image adaptive processing unit intended for processing the input information image in accordance with the algorithm developed in advance for recognizing information images; a comparison unit intended for comparing the input information image being processed in the information image adaptive processing unit with the model image, and for estimating discrepancies between the compared images; a model pattern correcting unit intended for correcting the pattern image by signals from the comparison unit. The drawbacks of this system are the same as those of the method being implemented by this system.
The task of this invention is to develop a method and a system for adaptive recognition of information images which could optimize the procedure for choosing the most exact and fast algorithms for processing a large flow of information in accordance with the type of the problem to be solved and the nature of the data being processed. In other words, this invention is aimed at solving the problem of accelerating a search for adequate processing and decision algorithms to make such work automatic. At the first stage, the system in which this invention has been implemented is necessary to develop and create automatic recognizing devices, and then can itself be included as a component part into an adaptive self-tuned recognizing system which is automatically aligned to changed operating conditions.
Thus, the technical result being reached is in increasing a recognition accuracy for the wider class of objects.
To solve this problem and reach the required technical result, a method is claimed for adaptive recognition of information images, comprising steps of: processing each input information image in accordance with a recognition algorithm; comparing the input information image being processed with a model image in order to estimate discrepancies between the compared images; comparing at least one estimate of the obtained discrepancies with at least one predetermined threshold; when said estimates exceed the predetermined thresholds, correcting the recognition algorithm, and repeating the previous operations using said recognition algorithm for the same input information image, wherein, in accordance with the present invention, said method further comprises steps of: forming models of probable information images; for the probable information images, creating the recognition algorithms based on the formed models; processing the information image in accordance with one of the created recognition algorithms; performing, after said processing step, a model recovering of the processed information image in accordance with the recovery algorithm which is inverse of the utilized recognition algorithm; the step of comparing the input information image being processed is performed with the recovered model image, calculating coincidence estimates, and then performing said step of comparing the estimates with the predetermined thresholds; when the predetermined number of estimates do not exceed the predetermined thresholds, deciding on the recognition of the input information image; correcting the recognition algorithm when the predetermined number of estimates exceed the predetermined thresholds, and then repeating, using the corrected algorithm for the same input information image, the steps of processing the information image, recovering the model image, comparing those images, and comparing the obtained estimates; repeating the previous step until deciding on the recognition of the input information image, or deciding on the absence of possibility for recognizing this image on the basis of an available set of algorithms.
In so doing, the step of processing each input information image includes at least steps of encoding and normalizing said input information image, and in the step of encoding, a step of at least rough encoding and exact encoding for the input information image can be performed.
It is distinctive that the step of model recovering is performed in the same stages as the step of processing the input information image but in inverse order with respect to the order in said step of processing the input information image.
In this case, the step of comparing the input information image being processed with the recovered model image is performed for all processing stages and corresponding recovery stages.
Another distinction is that the recovered model image is displayed for visual perception by an operator.
The step of forming models of probable information images is performed either in advance or during the mentioned comparisons, and the step of creating the algorithms for recognizing the probable information images is performed either in advance or during the mentioned comparisons.
Finally, the step of correcting the recognition algorithm is performed either by defining more precisely its parameters or its operation sequence, or by replacing this algorithm.
To solve the same problem and reach the same technical result, a system is claimed for adaptive recognition of information images, comprising: an adaptive information image processing unit intended for processing an input information image in accordance with a recognition algorithm; a comparison unit intended for comparing the input information image being processed in the adaptive information image processing unit with a model image, and for estimating discrepancies between the compared images; a model image correction unit intended for correcting the model image by signals from the comparison unit, wherein, in accordance with the present invention, said system further comprising: a memory unit intended for storing the information image processed by the information image adaptive processing unit at least until finishing the image recognition procedure; a processed information image model recovering unit intended for recovering the processed information image in accordance with a recovery algorithm which is inverse of the recognition algorithm utilized in the information image adaptive processing unit; and a decision unit intended for comparing at least one of the discrepancy estimates obtained in the comparison unit with at least one predetermined threshold, and for deciding in dependence on a result of that comparison; the information image adaptive processing unit being intended for processing the input information image in accordance with any of the created algorithms for recognizing the probable information images; the comparison unit being intended for comparing the input information image with the model image recovered by the processed information image model recovering unit, and for estimating discrepancies between the compared images; and the model image correction unit being made in the form of a recognition algorithm correction unit and intended for correcting the next recognition algorithm by a signal from the decision unit in the case when the discrepancy estimates from the comparison unit exceed corresponding thresholds and for realigning the information image adaptive processing unit and the information image model recovering unit in accordance with the corrected algorithms.
In so doing, the information image adaptive processing unit can comprise at least an encoding subunit and a feature normalizing subunit, the encoding subunit can consist of a rough encoding subblock and an exact encoding subblock.
The system can be equipped with a first and second demonstration displays for displaying the input information image and the recovered model image, respectively.
A specific feature of the system is in that it can further comprise: an additional information image adaptive processing unit intended for processing the input information image in accordance with any of the created algorithms for recognizing a probable noise and interference; an additional processed information image model recovering unit intended for recovering the processed information image in accordance with a recovering algorithm that is inverse of the recognition algorithm utilized in the additional information image adaptive processing unit; an additional comparison unit intended for comparing the input information image being processed in the additional information image adaptive processing unit with the model image recovered by the additional processed information image model recovering unit, and for estimating discrepancies between the compared images; an additional decision unit intended for comparing at least one of the discrepancy estimates obtained in the additional comparison unit with at least one additional predetermined threshold, and for deciding in dependence on a result of that comparison; the memory unit is also intended for storing the information image processed by the additional information image adaptive processing unit at least until finishing the image recognition procedure; the recognition algorithm correcting unit is also intended for correcting the next algorithm for recognizing a probable noise and interference by a signal from the additional decision unit in the case when the discrepancy estimates from the additional comparison unit exceed corresponding thresholds, and for realigning the additional information image adaptive processing unit and the additional information image model recovering unit in accordance with the corrected algorithms.
In this case, each pair of main and additional units having the same name can be made in the form of one common unit having the same name, and the recognition algorithm correcting unit can be also intended for alternating the probable information image and probable noise and interference recognition processes in each common unit.
No objects of the same purpose, containing the corresponding features of the objects described above, have been revealed in the existing engineering level. This makes it possible to consider the claimed method and system new.
From the background of the invention no sources of information are known comprising the whole set of the essential features of the described objects. This allows to consider the claimed method and system novel.
From the background of the invention no sources of information are also known where the set of the features have been described which could distinguish the claimed objects from their nearest analogues. This allows to consider the claimed method and system having the inventive step.